基于形状统计模型的多类目标自动识别方法

孙显; 王宏琦; 杨志峰

doi:10.3724/SP.J.1146.2008.01422

基于形状统计模型的多类目标自动识别方法

doi: 10.3724/SP.J.1146.2008.01422

基金项目:

国家自然科学基金(40871209)，国家863计划项目(2006AA12Z149)和中国科学院电子学研究所青年创新基金资助课题

计量
- 文章访问数: 3321
- HTML全文浏览量: 103
- PDF下载量: 739
- 被引次数: 0
出版历程
- 收稿日期: 2008-11-03
- 修回日期: 2009-06-25
- 刊出日期: 2009-11-19

Automatic Multi-categorical Objects Recognition Using Shape Statistical Models

摘要

摘要: 形状是人类视觉系统分析和识别目标的基础。针对现有方法的不足，该文提出了一种新的基于形状统计模型的多类目标自动识别方法。该模型定义形状基元对作为特征描述子，从样本图像中抽取典型基元对，聚类量化后组成形状字典。然后综合分析各类信息，通过无监督学习来统计目标的特征分布状况，构建类别形状模型。快速定位目标区域并辨识对象类别后，可结合图像分割获取精确形状。实验结果表明，该方法能准确、高效地提取多种类型和复杂结构的目标，较好解决了噪声干扰、旋转侧偏等问题，具有较强的实用价值。
- 图像处理; 目标识别; 形状统计模型; 无监督学习
Abstract: Contour features are powerful cues for human vision system to analyze and identify objects. A new method for automatic multi-categorical objects recognition using shape statistical models is proposed to improve the disadvantages existing in most of the relative methods. This method defines firstly the shape base pairs as feature descriptors, and extracts typical shape base pairs from sample images to build a feature codebook. Then, unsupervised learning is performed to calculate the feature distribution and design class-specific shape models. After detecting the regions and determining the categories quickly, segmentation could be applied to obtain the precise outlines. Experimental results demonstrate that proposed method can achieve high efficiency and accuracy in extracting manifold and complicated objects, and resolve the problems of noise disturbance, rotations at a certain extent.

HTML全文

参考文献(1)

Leibe B, Leonardis A, and Schiele B. Robust object detectionwith interleaved categorization and segmentation [J].International Journal of Computer Vision Special Issue onLearning for Recognition and Recognition for Learning, 2008,77(1): 259-289.[2]Felzenszwalb P F and Schwartz J D. Hierarchical matching ofdeformable shapes [C]. Proceedings of IEEE Conference onComputer Vision and Pattern Recognition, Minnesota, USA,2007: 1-8.[3]Zhang X Q, Guo M M, and Tang Y. A new geometric featureshape descriptor [J]. Computer Engineering and Applications,2007, 43(29): 90-92.[4]Fergus R, Perona P, and Zisserman A. Weakly supervisedscale-invariant learning of models for visual recognition [J].International Journal of Computer Vision.2007, 71(3):273-303[5]Opelt A, Pinz A, and Zisserman A. Incremental learning ofobject detectors using a visual shape alphabet [C].Proceedings of IEEE Conference on Computer Vision andPattern Recognition, New York, USA, 2006: 3-10.[6]Shotton J, Blake A, and Cipolla R. Multi-scale categoricalobject recognition using contour fragments [J].IEEETransactions on Pattern Analysis and Machine Intelligence.2008, 30(7):1270-1281[7]Mahamud S, Williams L R, Thornber K, and Xu K.Segmentation of multiple salient closed contours from realimages [J].IEEE Transactions on Pattern Analysis andMachine Intelligence.2003, 25(4):433-444[8]Kumar S. Models for learning spatial interactions in naturalimages for context-based classification [D]. [Ph.D.dissertation], The Robotics Institute, Carnegie MellonUniversity, 2005.[9]Wu Tao, Ding Xiao-qing, and Wang Sheng-jin. Videotracking using improved chamfer matching and particle filter[C]. Proceedings of IEEE Conference on ComputationalIntelligence and Multimedia Applications, Sivakasi, TamilNadu, 2007: 169-173.Maitre H, Kyrgyzov I, and Campedel M. Kernel mdl todetermine the number of clusters [C]. Proceedings of IEEEConference on Machine Learning and Data Mining in PatternRecognition, Leipzig, Germany, 2007: 203-217.[10]Hofmann T. Unsupervised learning by probabilistic latentsemantic analysis [J].Machine Learning.2001, 42(2):177-196[11]Dempster A, Laird N, and Rubin D. Maximum likelihoodform incomplete data via the EM algorithm [J]. Journal of theRoyal Statistical Scotiety, 1977, 39(1): 1-38.[12]Lempitsky V, Rother C, and Blake A. LogCut-efficient graphcut optimization for markov random fields [C]. Proceedings ofIEEE Conference on Computer Vision, Rio de Janeiro, Brazil,2007: 1-8.

施引文献

资源附件(0)

访问统计